SEAR enhances 3D reconstruction by integrating RGB and thermal imaging, optimizing visual geometric transformers for practical applications.

HORNet is a lightweight frame selection policy that significantly reduces video processing time and improves answer quality for vision-language models in video question answering tasks.

An AI assistant that automates product copywriting and provides real-time Q&A for live streamers, boosting sales and engagement.

A closed-loop framework that uses a VLM critic to steer text-to-image generation for improved spatial accuracy and state-of-the-art results.

Leverage implicit 3D knowledge from video generation models to enhance multimodal LLMs for spatial reasoning and embodied tasks.

A novel training framework for 3D Gaussian Splatting that enables continuous level-of-detail rendering from a single model without sacrificing quality, offering a smooth speed-quality trade-off.

This research provides a mechanistic understanding of how Vision-Language-Action models generate robot actions, revealing key insights into their internal representations and offering tools for interactive exploration.

A novel discrete diffusion model for high-dimensional visual generation that unifies understanding and generation tasks, with code available.

A unified framework for stable and efficient articulated 3D object reconstruction from single images, outperforming existing methods in accuracy and speed.

A benchmark and mitigation strategies for building more trustworthy embodied navigation agents that can withstand real-world corruptions.

A novel diffusion-based tokenizer bridges semantic understanding and precise kinematic control for realistic motion generation.

A generative inverse rendering method that disentangles object materials and illumination from a single image, enabling realistic scene reconstruction.

A new benchmark for evaluating LLM financial reasoning, integrating company fundamentals and trading signals, with available code and a clear market need.

A novel method and dataset for high-quality video object removal that also eliminates visual effects like shadows and reflections.

Develops online learning algorithms that use ranking feedback instead of numerical utilities, enabling applications in human-in-the-loop systems and privacy-sensitive scenarios, with demonstrated effectiveness in LLM routing.

A compact, high-performance LLM with strong reasoning and agentic capabilities, trained using novel cascade reinforcement learning and distillation techniques, with open weights and training data.

DriveTok provides a unified 3D scene tokenization method for efficient multi-view reconstruction and understanding in autonomous driving.

A novel vector field reshaping strategy for diffusion models significantly improves generative segmentation performance by addressing gradient vanishing and class separation issues.

A new benchmark and dataset for evaluating large language models on long-form audio-visual understanding, addressing a critical gap in current AI capabilities.

Generate semantically grounded, part-aware 3D objects from text by jointly modeling part geometry, appearance, and inter-part relationships.

This research proposes state space models as a more efficient and competitive alternative to Transformer-based vision encoders for large vision-language models, offering improved performance and robustness at a smaller scale.

A novel autoencoder architecture that significantly improves both image generation and editing quality by better aligning latent representations with pretrained visual models.

A prompt-tuning method that improves vision-language model performance by addressing framing-induced visual attention biases.

FASTER enables real-time responsiveness for vision-language-action models in physical world applications by optimizing action sampling and reaction latency.

A novel framework for autonomous driving perception that leverages 3D Gaussian Splatting to learn geometrically precise Bird's-Eye-View representations, outperforming existing methods.

A scalable multi-agent critic framework that significantly improves GUI agent robustness and performance by decomposing trajectories into verifiable milestones for accurate reward generation.

Develops novel membership inference attacks to quantify and improve the privacy of synthetic tabular data generated by diffusion models.

A probabilistic few-shot acoustic synthesis method that generates spatially continuous sound rendering for immersive environments with minimal data.

A new benchmark and evaluation framework for GPU kernel optimization that measures performance against hardware limits, enabling faster development of efficient AI models.

A decentralized MPC framework with CBF constraints for safe, real-time multi-robot quadrupedal locomotion, demonstrated on hardware.

A framework for trustworthy coronary angiography analysis that uses preference-aligned perception and RL-based reasoning to achieve topologically coherent vessel segmentation and reliable stenosis detection.

A probabilistic agentic framework that grounds complex natural language commands into actionable robot decisions in 3D space, outperforming existing VLM approaches.

A GPU-accelerated framework for combinatorial optimization that uses LLMs to convert natural language problem descriptions into executable solver code.

A training-free framework that uses adaptive prompt blending to improve the accuracy and fidelity of text-to-image generation for rare concepts and complex edits.

A training-free framework that deterministically schedules prompts for text-to-image models to generate rare compositional concepts with precise control.

A novel reinforcement learning framework to significantly improve low-resource language translation quality by optimizing subword segmentation and data imbalance during training.

A novel decoding framework for discrete diffusion models that enhances text generation diversity with minimal computational overhead.

A novel method for continual learning that adapts pre-trained models to new tasks with improved stability and expressivity, outperforming existing state-of-the-art.

A framework to debias LLMs by enforcing structural invariance in their internal graph representations, reducing bias without sacrificing performance.

SHAPCA provides consistent and interpretable explanations for machine learning models on spectroscopic data, enabling trust and adoption in critical applications.

A zero-shot embodied exploration framework using 3D Gaussian Splatting for persistent spatial memory and enhanced VLM reasoning.

An adaptive stock prediction system that uses autoencoders and reinforcement learning to dynamically adjust to market volatility, outperforming baselines.

An open-source, low-cost social robot platform designed for reproducible research and real-world deployment.

Develops a novel multimodal attack framework to identify and exploit vulnerabilities in spoken language models, enabling more robust AI safety solutions.

A novel generative image classification approach that leverages any-order autoregressive models to achieve state-of-the-art performance with significantly improved efficiency.

A framework for generating high-fidelity, instance-level 3D scene textures from reference images, enabling precise appearance editing.

A federated learning system that uses neuron-wise analysis and adaptive client rating to defend against targeted label-flipping attacks in road condition classification, ensuring transportation safety.

LuMamba is a highly efficient, topology-invariant EEG foundation model pre-trained on 21,000 hours of data, achieving state-of-the-art performance on multiple diagnostic tasks with significantly reduced computational cost.

A vision-language model and dataset for understanding traffic anomalies in real-world roundabout videos.

A dual-path RAG framework that significantly enhances multilingual multi-hop question answering accuracy by leveraging parallel query translation and bilingual retrieval.

Develop a framework to steer and audit the safety judgments of vision-language models by identifying and exploiting their reliance on semantic cues, addressing a critical vulnerability in deployed multimodal systems.

A new multi-modal dataset and network for detecting small building changes in remote sensing imagery, outperforming existing methods.

A real-time SLAM system that robustly handles dynamic environments by estimating per-pixel uncertainty, outperforming existing methods in cluttered and moving scenes.

A conditional neural solver for the Multi-Objective Multiple Traveling Salesman Problem that enables robotic teams to optimize competing objectives like cost and time, with demonstrated real-world applicability.

Augment robot simulators with accurate fire dynamics for realistic training and evaluation of firefighting robots.

An agentic LLM framework to automate and scale sign language annotation and dataset curation, overcoming the limitations of manual methods and gloss-level analysis.

A framework for efficient and accurate proactive streaming video understanding that decouples semantic understanding from perception.

SwiftTailor generates realistic 3D garments 10x faster than existing methods by unifying sewing pattern prediction and geometry-based mesh synthesis.

A new metric to accurately measure and identify 3D spatial geometric inconsistencies in generated videos, addressing limitations of current evaluation methods.

A framework for LLMs to generate scientifically grounded ideas by learning explicit reasoning processes from research motivations.

A neural network approach to autoregressive time series estimation that is significantly faster and more robust than traditional methods, offering interpretable coefficients.

TerraScope is a pixel-grounded visual reasoning model for Earth observation that enables precise geospatial analysis across multiple modalities and time points.

A diffusion model that uses image-derived priors to precisely remove reflections from single images, improving visual quality and detail.

ATG-MoE enhances robotic assembly skill learning by utilizing an efficient trajectory generation technique.

A zero-shot debiasing framework for vision-language models that disentangles bias from semantic information in a sparse latent space to improve fairness without sacrificing performance.

A novel approach to enable Multi-modal Large Language Models to perform segmentation directly, eliminating the need for external decoders and improving feature precision.

A black-box system to monitor and detect behavioral drift in LLM endpoints, ensuring AI application consistency.

This research introduces a new multilingual benchmark and evaluation methodology to precisely diagnose and improve temporal reasoning capabilities in Large Language Models, particularly for low-resource languages and diverse calendar systems.

A framework for generalized 3D hand-object pose estimation that handles occlusion by integrating semantic knowledge and hand priors.

A novel RAG framework that rewrites queries to retrieve decision-relevant evidence, improving accuracy on complex question-answering tasks.

AgentDS benchmarks human-AI collaboration in domain-specific data science, revealing the critical role of human expertise and guiding the development of next-generation AI agents.

A lightweight fine-tuning method for diffusion models that significantly reduces data and computational requirements while outperforming state-of-the-art.

A GAN-based simulation framework to quantify and mitigate racial bias in predictive policing systems, with publicly available code and data.

A framework for training humanoid robots to traverse complex terrains with natural gaits using reference gait priors and self-supervised terrain estimation.

A Bayesian model for drug discovery that uses variable selection to improve prediction accuracy and identify clinically meaningful drug-disease associations.

Deploying explainable and fair AI for anomaly detection in distributed power plants to ensure operational continuity and reduce maintenance costs.

A novel reinforcement learning technique that bootstraps learning with demonstrations to improve exploration efficiency and reasoning pattern acquisition for complex tasks.

A zero-shot panoramic depth estimation framework that leverages 3D consistency from foundation models to provide accurate geometry-aware depth maps without training.

An unsupervised contrastive learning approach for efficient and robust 3D shape matching that outperforms state-of-the-art methods.

Lightweight model predictive control for spacecraft attitude synchronization, optimized for real-time onboard execution in resource-constrained New Space missions.

A fast, category-agnostic framework using Gaussian Splatting to reconstruct realistic 3D hand-object interactions from RGB videos.

A safety-guaranteed imitation learning framework for spacecraft close proximity control that significantly reduces online computation while maintaining expert-level performance.

A privacy-preserving framework enabling cross-model inference between independent language models using secure linear alignment and homomorphic encryption.

Generate pathology-aware synthetic PET scans from MRI to improve neurodegenerative disease diagnosis, achieving near-actual PET performance with a 4% improvement over MRI.

A benchmark and training corpus to significantly improve the multi-hop spatial reasoning and visual grounding capabilities of vision-language models for embodied agents.

A framework that enhances visual in-context learning by intelligently fusing and aligning multiple visual prompts, improving performance across various vision tasks.

This research introduces a new dataset and training methodology to enable large language models to precisely derive and reason over complex mathematical objects, improving STEM application capabilities.

DriftGuard is a federated learning framework that efficiently adapts to asynchronous data drift by separating global and local parameters, reducing retraining costs by up to 83% while maintaining high accuracy.

A semantic-augmented DRL framework that leverages LLMs to improve UAV-aided vehicular network connectivity and efficiency.

RewardFlow enables more efficient and effective reinforcement learning for large language models by providing fine-grained, state-level rewards derived from the topology of reasoning trajectories.

A motion-centric video understanding model that makes object trajectories explicit and verifiable, enhancing spatial-temporal reasoning.

Develop interpretable foundation models for retinal imaging that provide faithful local and global explanations, outperforming state-of-the-art models with greater parameter efficiency.

FEDB dramatically reduces the performance overhead of confidential databases by removing cryptographic operations from the critical path, enabling faster and more efficient secure data queries.

A novel denoising network for rapid, high-resolution transmission electron microscopy imaging, enabling atomic-scale observation of material nucleation dynamics.

A formal specification and reference implementation for secure, auditable, and policy-compliant autonomous agent actions in B2B environments.

A multi-stage framework for detecting human values in noisy Russian social media text, with publicly released models and benchmark results.

A scalable API service for RL training of multi-turn LLM agents, simplifying rollout orchestration and providing standardized sandbox environments.

An AI agent that generates novel and valuable mathematical research problems in differential geometry, validated by human experts.

Automate robotic simulation and trajectory synthesis using video generation priors to create diverse, physically grounded environments and expert trajectories from natural language.

A novel policy optimization method for diffusion large language models that significantly reduces training costs and improves performance on instruction following and reasoning tasks.

VesselTok enables efficient and generative modeling of complex 3D anatomical structures like blood vessels and airways using novel tokenized latent representations.

Perceptio enhances vision-language models with explicit 2D and 3D spatial reasoning by generating semantic segmentation and depth tokens directly within the autoregressive sequence.

A robust LLM watermarking framework that anchors ownership signals into a stable functional subspace, outperforming state-of-the-art against common model attacks.

A novel approach to disentangle and quantify uncertainty in medical image segmentation, improving reliability for safety-critical applications.

Weaver is a greybox fuzzing framework that finds critical vulnerabilities at the JavaScript-WebAssembly boundary in web engines.

A 32B parameter LLM optimized for multi-step reasoning and long-context understanding in Korean enterprise environments, with state-of-the-art performance on domain-specific benchmarks.

A visual-tactile imitation learning system for robots to reliably trace and manipulate deformable objects, improving generalization across object types.

A diffusion-based framework that uses point cloud priors to generate geometry-controllable 3D assets and scenes with superior quality and fidelity.

Automate LLM post-training pipeline configuration to achieve state-of-the-art performance with significantly reduced computational cost.

A framework for precisely and robustly removing unwanted concepts from text-to-image models using diverse prompt representations.

A deep learning VAE-GAN model enhances reservoir simulation by improving parameterization for better history matching and geological realism.

This research reveals and quantifies implicit grading bias in LLMs, providing a critical tool for developing fairer automated assessment systems.

A probability calibration method for source-free domain adaptation that preserves source knowledge and reduces noise overfitting in computer vision tasks.

A red-teaming framework to evaluate the real-world security of autonomous web agents against network-layer threats.

A novel transformer architecture inspired by game theory and statistical physics to improve token dependency modeling and achieve state-of-the-art performance on NLP tasks.

A hybrid CNN-SSM architecture for domain adaptive object detection that aligns global and local features efficiently.

A simplified reinforcement learning approach that enhances LLM reasoning and mathematical abilities with less complexity than existing methods.

A weakly supervised AI that generates faithful and plausible natural language explanations for chest X-ray classifications, improving diagnostic accuracy and adapting to different audiences.

Accelerate video generation by up to 1.92x and reduce memory by 3.32x using inference-time mixed-precision quantization and temporal caching for diffusion transformers.

This research quantifies and exploits confirmation bias in LLM-assisted code review, demonstrating a significant vulnerability in AI security tools that can be mitigated through prompt engineering and metadata redaction.

EdgeCrafter enables high-performance dense prediction tasks like object detection and instance segmentation on resource-constrained edge devices using compact Vision Transformers through task-specialized distillation.

CausalRM enables cost-effective LLM alignment by learning reward models from readily available observational user feedback, overcoming noise and bias to significantly improve downstream performance.

A neuro-symbolic framework that uses structured knowledge to bridge the simulation-to-reality gap in image translation, enabling data-efficient and generalizable zero-shot transfer.

A multi-agent framework that coordinates LLM memory cycles for improved long-horizon interaction and self-evolving memory repair.

A novel multi-channel time series forecasting architecture that uses sparse attention and multimodal fusion to achieve competitive accuracy with lower computational cost.

A pretraining framework that distills knowledge from existing time series foundation models to create a unified encoder for scientific time series data, improving representation learning.

Enhance agentic decision-making in complex, long-horizon tasks by modulating rewards with hindsight information to improve credit assignment.

We've created the first benchmark and analysis for audio pun understanding in large audio-language models, revealing critical performance gaps and providing insights for future development.

A novel attention mechanism inspired by cognitive science that offers linear-time scaling and enhanced representational power for multimodal AI tasks.

A novel method to improve the topological accuracy of image segmentation, enhancing reliability for downstream analysis.

A neural network learns a continuous distance field in robot configuration space for safe, zero-shot motion planning in complex environments.

A framework and benchmark for teaching multimodal LLMs to strategically use visual aids for geometric reasoning, improving performance by 3.51%.

A novel framework for speech anti-spoofing that significantly improves model robustness across diverse datasets by extracting domain-invariant features, reducing error rates by 20%.

A novel training method for vision-language models that significantly improves reasoning accuracy and conciseness by adaptively weighting training segments, reducing training time and cost.

A parameter-efficient semi-supervised learning framework for medical ultrasound image segmentation that leverages multiscale patch mixing and frequency domain contrastive learning to achieve state-of-the-art performance with limited labeled data.

A new LLM-powered evaluation framework for PDF table extraction that significantly outperforms existing metrics, providing practical guidance for parser selection.

A diffusion-based face restoration method that bridges the age gap in identity-preserving image enhancement, enabling historical photo restoration with cross-age references.

PhysVideo generates physically plausible videos by leveraging cross-view geometry and physics-aware attention, addressing a key limitation in current video generation.

A framework for optimizing LLM fine-tuning budgets to balance safety, reduce over-refusal, and improve instruction following.

A training-free sparse attention framework for significantly faster video generation with minimal quality loss.

D-Mem is a dual-process memory system for LLM agents that combines fast vector retrieval with high-fidelity deliberation to improve long-horizon reasoning without significant computational overhead.

A framework for identifying terrestrial analogs of extreme deep-sea environments using topographic similarity, enabling cost-effective biological research.

A fine-grained benchmark and evaluation framework to diagnose and improve AI-generated video detection capabilities of Large Vision-Language Models.

A training-free framework for zero-shot object-goal navigation that uses LLMs to reason over a tree of spatial paths for efficient exploration.

A new open-source dataset and pretrained model for high-quality text-to-motion generation, addressing data limitations and improving generalization.

A reinforcement learning framework that trains LLMs to iteratively refine their own context for improved performance on new tasks, outperforming existing methods.

A diagnostic environment to rigorously evaluate and improve the reasoning-action coupling of tool-augmented LLMs.

Develops an intelligent digital twin system that detects and discriminates cyber-attacks in industrial control systems, enabling safe operation without costly shutdowns.

A benchmark and baselines for unsupervised discovery of phoneme inventories from speech, enabling better understanding and processing of unseen languages.

A multimodal AI that explains its humanitarian crisis classifications by transferring rationale between text and images, improving accuracy and reducing annotation needs.

AutORAN automates the creation of cellular network applications using natural language, reducing development time from months to minutes.

Develop a multilingual embedding tool with superior performance on language benchmarks.

A customizable, low-cost, 3D-printable tendon-actuated robot with a tapered backbone for compliant inspection and manipulation tasks.

An AI system using guided vision-language models for coordinating multi-actor human-robot interactions.

This research explores and quantifies linguistic stereotypes in LLM outputs across different dialects, demonstrating the effectiveness of prompt engineering and multi-agent architectures for bias mitigation.

This research develops a cost-aware framework to identify and quantify the vulnerabilities of phishing detectors to feature manipulation, revealing that adversarial robustness is driven by feature economics, not model complexity.

Develop a visuo-tactile robotic system for improved contact-rich manipulation tasks.

A novel signature method for time series analysis that offers richer memory dynamics and improved expressivity over existing methods.

Create tools for interpreting and exploring features in VLA models using sparse autoencoders.

A new architectural framework for LLMs that enforces reasoning integrity through explicit control layers, significantly reducing hallucination and improving reliability.

A dynamic mixed-precision quantization framework for efficient Mixture-of-Experts inference on edge devices.

A new annotated dataset and benchmarking for identifying patient health literacy from clinical notes to improve patient outcomes.

A new AI framework improving robotic movement through state-of-the-art dynamics-grounded learning.

A federated learning framework for multi-modal data that uses latent-space alignment to improve efficiency and robustness.

A lightweight cryptographic protocol for verifying AI model inference in cloud services, reducing proving times from minutes to milliseconds.

A minimalist approach to fingerprint enhancement that achieves state-of-the-art results with simpler, more effective methods.

A new evaluation suite for LLM-powered survey simulation that captures both ranking and distribution alignment, enabling more meaningful and comparable assessments.

A platform for distributed Turing Tests where humans and LLMs interact and judge each other to monitor AI evolution.

A structured prompting framework that significantly reduces follow-up prompts and improves AI intent alignment for ambiguous tasks.

Accelerate LLM agent task completion by speculatively executing tools based on recurring patterns and data dependencies.

Leveraging LLM self-reports to track and steer internal emotive states for safer and more interpretable AI.

A framework for accessible text generation that integrates human feedback to improve LLM-based simplification and evaluation.

An LLM-aided framework that translates natural language into structured constraints for joint wireless network optimization, outperforming traditional methods.

Leverage early ocular and cardiac signals to predict user performance in interactive systems, enabling proactive interventions.

Predict audience engagement and vocal attractiveness in learning videos using only speaker expressiveness, enabling scalable and privacy-preserving affective computing.

Develops and benchmarks neural network models for detecting AI-generated text, outperforming existing commercial tools on specific datasets.

A novel polynomial kernel for Gaussian Splatting that improves performance and maintains dataset compatibility for 3D reconstruction.

A novel off-policy learning method for contextual bandits that optimizes item allocation under limited supply constraints, outperforming existing methods in real-world scenarios.

Scalable sparse commodity recommendation for industrial applications using orthogonal constrained projection.

This research empirically compares catastrophic forgetting mitigation strategies for continual intent classification, finding replay-based methods combined with architectural choices are crucial for robust adaptation.

A method to improve treatment effect estimation from randomized trials by aligning data from observational studies with covariate mismatches.

This paper evaluates the performance of ChaCha20-Poly1305 encryption for IoT and ICS devices, demonstrating its feasibility within strict latency requirements.

Optimizes compute allocation for language model pretraining and specialization to improve performance across diverse benchmarks.

LLMs can generate better analogies than humans by adhering to relational constraints, suggesting a new approach to analogical reasoning.

An adaptive algorithm for nonlinear data assimilation that automatically balances model complexity using P-splines and an information criterion.

This research revisits and benchmarks a deep learning anomaly detection model against a simpler PCA-based approach, highlighting the impact of evaluation methodology on perceived performance.

A novel reinforcement learning exploration method that improves agent visitation of features within trajectories by optimizing future state-action distributions.

A diagnostic study reveals that the shape of uncertainty dynamics in LLM reasoning steps, specifically entropy trajectory monotonicity, can predict reasoning reliability more effectively than aggregate uncertainty measures.

A tool to verify the security of hardware designs generated by High-Level Synthesis, preventing false positives and detecting HLS-induced flaws.

An optimized ensemble framework for financial loan default prediction that dynamically weights models based on performance to improve accuracy and interpretability.

A progressive training pipeline for explainable, bilingual (English-Hindi) knowledge-grounded dialogue systems that eliminates factual hallucinations.

A new framework for evaluating human-AI team readiness beyond simple accuracy, focusing on safe and effective collaboration.

A study proposes personalized, domain-specific language learning lessons generated by LLMs to bridge the gap to professional fluency, building on existing general lesson effectiveness.

Develops a passive folding mechanism for air-dropped sensors to enable low-cost, wide-area environmental monitoring.

A post-processing framework that enhances fairness in machine learning predictions without altering model internals, applicable across diverse tasks and fairness definitions.

A framework to guide practitioners in selecting and evaluating differentially private graph release methods, addressing privacy-utility trade-offs.

A framework for generating instruction-aware co-speech gestures for humanoid robots in educational settings.

A deep learning method for robust visual feature tracking in challenging environments, integrated into a VIO system.

A provenance-aware vulnerability analysis approach for Python applications that resolves vendored libraries to specific OS package versions or upstream releases to reduce false positives.

This research reveals a fundamental vulnerability in vertical federated learning label inference attacks and proposes a zero-overhead defense by adjusting model layers.

A novel statistical framework for detecting side-channel leakage in neural networks, offering improved sensitivity over traditional methods.

Develops a rigorous benchmark and evaluation framework for model-free policy optimization in complex, masked-action environments, revealing limitations of current methods.

A meta-learned system for rapid 3D surface reconstruction from satellite imagery using episodic priors.

Benchmarking CNNs against Transformers for abdominal organ segmentation reveals CNNs outperform on heterogeneous datasets, suggesting a focus on optimized CNNs for specific medical imaging tasks.

A principled method for designing per-instance noise schedules in diffusion models to improve generative quality, especially in low-step regimes.

This paper establishes the equivalence between prediction-powered inference and established survey sampling methods, offering a theoretical framework for integrating machine learning with statistical inference.

This paper analyzes implicit patterns in LLM-based binary vulnerability analysis to build more reliable systems.

This research redefines efficiency metrics for vision-language-action models, highlighting the disconnect between computational efficiency and real-world embodied performance on robotic platforms.

This paper investigates how to improve the reliability of reasoning language models by combining different uncertainty estimation techniques, showing that a hybrid approach offers significant gains.

This paper argues that Spectral GNNs for node classification are theoretically flawed and their performance gains are not due to spectral properties but rather message-passing dynamics.

This research explores how cross-domain analogies impact creativity in humans and LLMs, revealing differences in their ideation processes.

This paper provides a theoretical analysis of privacy and convergence in differentially private wireless federated learning, addressing open questions on privacy loss characterization and convergence guarantees for non-convex objectives.

Research into how LLM agents perceive security environments to enable privacy-preserving negotiations.

This paper develops theoretical regret bounds for competitive resource allocation in modular systems with endogenous costs, offering a formal justification for decentralized allocation strategies.

Develops a planning agent to evaluate difficulty in Tetris variants, offering insights for game design.

Develops unified algorithms for multi-dueling bandits that perform optimally in both stochastic and adversarial environments without prior knowledge.

A new NLP framework to analyze complex ideologies beyond the traditional left-right spectrum, enhancing social discourse analysis.

A theoretical framework for adaptive contextual bandit learning with semiparametric models and robust inference.

A conceptual framework for governing autonomous agents in business processes, focusing on alignment and operational autonomy.

A theoretical framework for approximating diffusion model transition probabilities by solving Fokker-Planck equations using Neural Galerkin Normalizing Flows.

This paper investigates corruption in multi-agent AI governance systems, finding that institutional design is a stronger driver of rule-breaking than model identity, suggesting a need for pre-deployment stress testing.

This research investigates how AI mediation in video communication impacts interpersonal trust and judgment confidence, finding a decline in both without affecting actual lie detection accuracy.

A theoretically complete and scalable algorithm for multi-agent path finding with asynchronous actions.

This research investigates why explicit cross-lingual alignment techniques often fail to improve downstream task performance, providing insights into optimizing alignment and fine-tuning strategies.

A theoretical framework to improve the generalization of any generative model using discriminators, with potential for new algorithms.

A co-designed social robot to assist young newcomers with language and cultural learning in community literacy programs.

This paper analyzes the security and deployability of IEEE 802.11az/bk secure Wi-Fi ranging, identifying vulnerabilities and providing guidelines for improvement.

A review of multimodal AI for computational pathology, focusing on representation learning and image compression for improved diagnostic accuracy.

A philosophical framework for governing increasingly capable synthetic minds, moving beyond tool-centric regulations.

This paper evaluates LLMs' strategic reasoning in game theory, highlighting limitations in counterfactual scenarios.

This paper provides theoretical guarantees for the accuracy of neural network solutions to partial differential equations by connecting residual error to solution-space error.

A computational framework for discovering hierarchical structure in sequential data through online inference.

This paper analyzes the effectiveness of UK cybersecurity reporting regulations using real-world incident data to inform policy.

Establishes theoretical lower bounds for linear classification problems, closing a gap in understanding computational complexity.

This paper reviews the integration of AI in judicial decision-making, highlighting concerns and research gaps in risk assessment tools and human-AI interaction.

A theoretical framework unifying generative modeling approaches by framing distribution transformation as an optimal stochastic bridge problem.

This paper introduces a new framework for understanding agent intentions within causal systems, enabling empirical detection and discovery of goals.

A unified taxonomy for deep learning-based multivariate time series anomaly detection to systematize research and identify trends.

This paper analyzes EU regulatory provisions for AI, focusing on security, privacy, and agentic AI to inform policymakers and developers on compliance and governance.

Provides theoretical guarantees for the Adam optimizer, improving understanding of deep learning training.

This paper explores how AI systems represent and reason about geography, highlighting potential biases and limitations in their understanding of spatial concepts.

This research explores student perspectives on AI ethics and social impact to inform future AI education.

This paper theoretically analyzes and improves a statistical method for approximating the Wasserstein distance, with no immediate product application indicated.

A framework to measure if AI amplifies human intelligence or makes us dependent on it, guiding sustainable AI system design.

This paper provides a theoretical analysis of No-U-Turn Sampler variants, offering convergence guarantees and mixing time results for Gaussian targets.

This paper introduces a complexity hierarchy for shuffles in card-based cryptographic protocols to evaluate their practical implementation.

A theoretical framework for understanding how the brain regulates stress and behavior under uncertainty by introducing 'Authority-Level Priors' to constrain hypothesis selection for control.

This volume collects research papers on the intersection of Artificial Intelligence and Theory of Mind.

Memento-Skills explores agent-based AI designing other agents, focusing on theoretical frameworks without product-ready applications.